Deep Learning for Binary Post-Earthquake Building Damage Detection from Multi-Temporal VHR Imagery

Rapid, reliable mapping of earthquake-induced building damage is essential for directing emergency response. This thesis tackles building-level damage detection from pre/post-event very-high-resolution satellite imagery, using the ESA Φ-Lab AI for Earthquake Response Challenge as context. The challenge targets automatic identification of damaged vs. undamaged buildings from image pairs and building polygons, leveraging historical Charter activations and multi-mission VHR data (e.g., Pléiades-1, WorldView-2/3, GeoEye-1, Kompsat-3/3A, Gaofen-2) to simulate real-world conditions. We present an end-to-end pipeline designed for three persistent difficulties: severe class imbalance, image misalignment, and geographic domain shift. Supervision is restricted to building footprints to limit label noise, and a pixel-level weighting term (computed from the ratio of valid negatives to positives) is integrated into the loss to avoid discarding scarce positives. To mitigate effective resolution differences across sensors, we apply pan-sharpening where panchromatic band is available, fusing the high-resolution pan band with lower-resolution multispectral bands to enhance spatial detail before learning. We adopt a Siamese architecture with a shared DINOv3-pretrained ConvNeXt encoder for the pre and post-event images; their embeddings are combined to predict building damage. Experiments on held-out areas and events assess classification performance and robustness of the final system. We discuss limitations (cross-sensor variability, footprint quality, and generalization across geographies) and outline practical next steps toward deployment, including multi-class severity estimation, domain adaptation. Overall, the work delivers a reproducible pipeline and actionable insights for scaling satellite-based, post-event damage mapping in real response scenarios.